Agent memory needs a shared format, argues LoreKit's founder

A blog post titled "Agent memory needs its OpenTelemetry moment" reached 72 points on Hacker News on 7 November 2026. Its author is Mads Thines of LoreKit. He makes the case that memory for coding agents has broken into formats that cannot read each other. He asks for a shared specification on the lines of the observability standard.

Thines opens with the shape he sees in one working day. A team runs four separate stores of agent memory at once. The first is an instruction file, usually called AGENTS.md, that lives in the repository. The second is an agent-managed file, which Anthropic's memory tool writes on its own. The third is hosted facts, which GitHub Copilot keeps in its own repository. The fourth is a knowledge graph behind an MCP server. He writes that "none of them can read the others". He adds that "the problem isn't a missing tool, it's a missing agreement".

The lock-in he names

"A lesson your CI agent learns stays in whatever wrote it," Thines writes. He adds that the editor agent cannot read it, and neither can the next tool a team tries. He draws the parallel with observability tools before 2019. Each vendor had its own agent. Traces did not cross tools.

OpenTelemetry solved that with a shared data model and semantic conventions. Any backend could read any trace. Thines wants the same shape for agent memory.

The four layers he proposes

The proposal has four layers. Layer one is a data model that defines what a memory record is. Layer two is semantic conventions for shared naming, so two tools use the same word for the same fact. Layer three is a protocol for live addressing, which he suggests should run over MCP. Layer four is a set of conformance tests. He does not want the group to pick one implementation.

He discloses his position in the post. "I build one of these tools, LoreKit," he writes. He calls the format "one proposal among several, with gaps of its own". That disclosure is up front in the piece.

For a team running more than one agent

The piece lands on a gap any team using several coding agents already feels. A fact learned inside one tool does not help the next. A team can act today in two small ways. It can keep its AGENTS.md as the source of truth and write to the other stores from it. It can also log where each agent is writing, so a review can read what the agents think they know.

Thines ends on a soft ask: the piece is an invitation to the group of vendors and users, not a demand on one of them.

Source

Agent memory needs its OpenTelemetry moment by Mads Thines, LoreKit. Discussion on Hacker News.